Are We Heading to an AI Bubble?

Appeared in Substack on November 7, 2025: https://anandanandalingam649613.substack.com/p/are-we-heading-to-an-ai-bubble?r=o7w77

In the roaring 1990s, digital technology was having a heady time. The Internet was becoming ubiquitous, telecommunications technology was making significant advances, there was a “land rush” to obtain wireless spectrum, and dot com companies were popping up like mushrooms in a damp wooded forest. Enormous amount of venture money poured into tech start-ups and well-established Wall Street firms fought each other to buy up the shares of the more mature information and communications companies. Between 1995 and 2000, the stock market experienced significant growth because of the excitement of dot com and tech companies. The Nasdaq Composite index rose close to 600 per cent between January 1995 and March 2020. During this period, close to $250 Billion of venture money was spent on technology. By 1999, 39% of all venture capital investments were going to Internet companies, many of them who were neither profitable nor with any revenue to show for their activities. That year, most of the 457 initial public offerings (IPOs) were related to Internet companies, followed by 91 in the first quarter of 2000 alone.

The dot com companies were on top of the world for about 5-6 years. Starting in March 2020, there was the dot com “crash” of historic proportion. The NASDAQ went from 5,048 on March 10, 2000, to 1,114 on October 9, 2002, a drop of 78 percent. During the dot-com crash, many online shopping companies like Pets.com, Webvan, and online search companies like Alta Vista and Lycos failed and shut down. Others, like Lastminute.com, MP3.com and PeopleSound were bought out. Larger companies like Amazon and Cisco Systems lost 80% of its stock value.

The bloodbath in technology companies was quite severe. Several communication companies, such as NorthPoint Communications, and Global Crossing, failed and shut down. WorldCom, one of the largest telecom companies at the time failed and was acquired by Verizon in 2006. Given the promise of wireless communication at that time, telecom companies competed vigorously with each other to buy spectrum. In Europe, this spectrum purchasing frenzy led to companies spending over $100 billion to acquire wireless spectrum and then were left with no finances to actually build a network. The mad scramble to acquire optical switches to speed up the telecom networks led to a slew of really bad decisions made by several big companies to buy optical start-ups who could not deliver. Many entrepreneurs made a lot of money even though their start-ups never produced anything of value even after being acquired by the likes of Lucent and Cisco.

Hank Lucas and I wrote a book called “Beware the Winners Curse” detailing many of the ill placed excitement during the 1990s and dissecting the reasons for this disastrous dot com and telecom tech period. Reasons for the fast rise of dot com and tech companies and their eventual (predictable) fall were unchecked speculation, irrational exuberance, and the widespread belief in the infallibility of the burgeoning internet sector. In fact, we provided compelling evidence for several causes of the dot com boom including hubris (invulnerability, optimism), herd mentality (fear of missing out), compensation systems of CEOs and investment firms (win or lose, money will be made) etc. We see all these forces acting out again in the excitement about Artificial Intelligence, both in developing the technology and in using it.

According to a recent article in The Times (UK) by Katie Prescott, citing Pitchbook, more than 40% of venture capital investment is currently going into just 10 AI-based companies. In 2024, it is estimated that OpenAI, Athropic and Databricks have absorbed more than $100 billion in venture capital investments while the total AI revenue was only $12 Billion. In the 2024 fiscal year, Databricks was valued at $100 billion with a revenue of $2.6 Billion and Anthropic was valued at $183 billion with a revenue of $3-$5 Billion. The revenue at OpenAI was around $4 billion. Outside of AI, the venture money is at record low. It is clear that there is a great deal of both herd mentality and hubris when it comes to venture investment in AI; no investor wants to miss out and hence are not doing the same rigorous due diligence as for other investments. So called experts of all variety and shades in academia, Wall Street, and consulting firms are extolling the amazing potential AI-based transformation that is about to sweep through the entire world. So much exuberance!

Why is all this money pouring in? There has always been excitement about having a computer-based system being able to substitute for human intelligence and do many tasks efficiently at speed and low cost. The promise of AI has been around for over 50 years but the progress was slow and with fits and starts. This excitement about AI was put on steroids when ChatGPT was released on November 30, 2022 and ordinary people could get extraordinary responses to simple queries. By January 2023, ChatGPT had become the fastest-growing consumer software application in history, gaining over 100 million users in two months. As of mid-2025, ChatGPT’s website is among the 5 most-visited websites globally, and has over 800 million active weekly users. It has been lauded as a revolutionary tool that could transform numerous professional fields.

Several leading-edge companies with massive push from the management consulting world have started examining AI for adaptation. An academic study by Brynjolfsson, Li and Raymond claims that AI could boost worker productivity by 14% and around 34% for new and low-skilled workers. PwC claims that AI could add $16 trillion to global GDP by 2030. McKinsey wrote several reports that provided very optimistic projections about how AI will transform industries and the U.S. economy. One of their reports estimated that AI could add $2.6-$4.4 trillion in annual value across corporations globally. They also “found” that 1% penetration of AI was associated with 14.2% increase in factor productivity. Another study of Japanese firms found that AI investments could lead to 2.4% increase in factor productivity. McKinsey also projected that generative AI could boost labor productivity by 0.1% to 0.6% per year through 2040. All these studies and projections were done before companies around the world had really incorporated AI in any significant way. This is how excitement and optimism is created that could lead to irrational exuberance about the future.

A McKinsey study reported that 92% of companies are investing in GenAI but only 1% of those already invested believe that their investments have reached “maturity”. Although 69% of the companies McKinsey surveyed had invested in GenAI in the past year, 47% of them say that the tools are getting developed too slowly. According to an MIT Sloan study, some manufacturing firms actually saw a drop in productivity when they first adopted AI. According to Observer, 80% of companies invested in AI are not seeing significant earnings gain. Reports from consulting firms are much more optimistic but, of course, they have an in-built conflict of interest: Consulting firms need to show that companies are not adopting AI correctly or too slowly but there are massive gains to be had if they listen to the management consulting advice. There are seemingly several cracks in the hype about the rapid expansion of AI in its different manifestations.

It is widely reported that LLM models, the leading edge of excitement in AI, are running out of learning data. There is talk now about using “synthetic” data rather than go after new data, i.e. using data that was generated by model implementations. To make matters worse, several groups including the New York Times are suing companies like OpenAI for using data from past articles without permission or compensation. Copyright and privacy laws will surely kick in when the different sectors of the economy realize that their intellectual property has been appropriated by AI companies without attribution or royalty payments. In early 2023, visual artists Sarah Andersen, Kelly McKernan, and Karla Ortiz filed a class-action lawsuit against Stability AI, Midjourney, and DeviantArt. They alleged that these companies used their copyrighted artworks without permission to train AI models like Stable Diffusion, leading to unauthorized reproductions of their styles and works. This lawsuit is notable as it represents a collective action by artists against AI companies, challenging the legality of using their works as training data. The AI art companies tried to have the case dismissed but a judge ruled in August 2024 that the could move forward. The final verdict will not be seen until the end of 2026, but this is not a good omen for AI companies on the hunt for new and expanded data. The outcome of these IP violation cases will not arrest the progress of AI but could slow it down and also add cost, how significant we don’t know at this stage.

Then there are litigations related to privacy violations by data hungry AI firms. According to the Stanford AI Index Report, AI-related privacy incidents surged by 56.4% in 2024. One case in point is Clearview Inc. which markets its tools mainly to law enforcement and says it has collected more than 60 billion images globally. It was recently sued by the Austrian government for privacy violations. The company had previously been found in breach of the GDPR by regulators in France, Greece, Italy, and the Netherlands in collecting and processing the data of millions of European citizens. The countries issued nearly 100 million euros ($116.62 million) in cumulative fines and reached a U.S. class-action settlement in March over its data-scraping practices. This is just one firm but given the precedence, it is likely that every time AI firms try to expand the collection of new data to keep their models dynamic and current, they will run into different laws around the world related to privacy and intellectual property rights. These litigations would be very costly and would likely slow down the adoption of AI in several sectors and applications.

Massive data centers are necessary for AI to be successfully deployed across several industries.  The response time of AI models needs to be really fast because the main strategic business value comes from being able to act in real time. For the response time to be very short, the computing power needed to process reams and reams of data has to be massive, hence the need for enormous number of servers. These servers require electricity for power and water for cooling. It is estimated that each ChatGPT search uses 10 times more electricity than a Google search. Training a single large AI model can use as much electricity as 100 U.S. homes use in one year. A single data center can use as much power as the City of Philadelphia. Some estimate that AI data centers in 2030 could consume 1,000 terawatt-hours of electricity annually which is more than the 2023 electricity consumption in all of the United Kingdom. Global data center electricity demand is projected to be nearly 1,000 Tera watt-hours in 2026, roughly 4% of global electricity use. AI data centers are typically clustered near fiber hubs (e.g. Northern Virginia, Frankfurt, Singapore and Dublin) which causes regional electricity grid stress. Data centers in Ireland started consuming over 18% of national electricity; new data center connections were temporarily paused in 2025. In the U.S., the current projections are that electricity demand from AI and data centers could reach 5-6% of total U.S. consumption by 2030.

Most AI data centers use evaporative cooling which is water intensive. In a March 2024 report, Forbes reported that a single conversation with ChatGPT would consume the equivalent of 1 liter of water to cool down the servers. A typical hyperscale data center can use 3-5 million gallons of water every day which is roughly what is consumed in a town of 30,000-50,000 people. In 2022, Microsoft reported that its global water consumption rose 34% largely attributed to AI data centers. Sooner or later, these data center projects will have to deal with public opposition and local and state regulations on water use and effluents. There is of course, talk about using alternate methods of cooling including immersing the data centers under water or using sea water cooling (Google in Hamina, Finland), air cooling and liquid immersion cooling (Microsoft in Phoenix, Arizona), each alternative costing enormous amount of capital investment. Google recently reported that over 20% of its data centers now use non-potable or reclaimed water for cooling. The bottom line is that implementing AI solutions will likely cost a lot more money than the speculative experts are admitting.

Although the dot com bubble burst in early 2000 and the effects were felt for several years, eventually several forward-thinking companies survived and thrived including Google, Amazon, Microsoft and Meta and several others like Expedia, Etsy, Twitter (now X) etc. The ultimate success of the “dot com” and technology companies took more than two decades. Many companies, especially exciting start-ups ended in the graveyard opf entrepreneurship.  In the long run, AI is here to stay and there will be very successful companies that create cutting-edge technology, and companies that integrate AI into business and the lives of people. Today one needs to be much more sober than all the hoopla one hears in the media and from the financial community. One would never believe listening to Wall Street analysts and Fortune-100 CEOs, not to mention management consulting firms and Ivory Tower academics, that the hurdles for implementing AI software, including building cost effective data centers, would slow down the adoption of AI systems.  

Not everyone lost when the dot com bubble burst. Many dot com managers and investment bankers made a lot of money. As the late MIT professor Charles Kindleberger put it “speculation tends to detach itself from valuable objects and turns into delusive ones. A larger and larger group of people seeks to become rich without a real understanding of the processors involved. Not surprisingly swindlers and catchpenny schemes flourish.” The share prices of AI companies might continue to climb for some time because of the greater fool theory. People who buy the shares of companies with doubtful business models will persuade themselves that there will always be someone, a greater fool, who will be willing to take the risk and buy their shares. If there is no greater fool than oneself, then the consequences could be dire. In the meantime, those who will make the most money will be consultants and financiers who are not directly involved in making sure that AI is good for business and society. Everyone else who are invested in AI stocks and are seeing these go up by leaps and bounds would surely be subject to the “winners curse” sooner than later.